Home / Infrastructure / Memgraph HA cluster reference

Memgraph HA Cluster Reference (orchgraph)

Topology at a glance

State on 2026-10-23

• All five containers are RUNNING in the Incus project default, with boot.autostart true. They run Archlinux images on the t9-btrfs pool.
• The prod env profile connects with KG_URI=neo4j://10.53.1.160:7691 and KG_DATABASE=orchkg.
• MAIN moves between the two data instances. The daily backup manifests from 2026-10-09 to 2026-10-22 record MAIN at 10.53.1.123 on six days and at 10.53.1.128 on eight. On the last run, 2026-10-22 03:00, MAIN was 10.53.1.123 (orchgraph-data-2).
• Node counts in the last backup manifest: memgraph 19,973 and orchkg 156,953. On 2026-10-09 they were 19,737 and 115,995.
• On 2026-10-21 a DeepChat session stopped the prod cluster. It was running again on 2026-10-22.

Writes auto-route to whichever data instance currently holds MAIN — the coordinators' Raft group decides that, and it can move on failover. Never hardcode a data-instance address for writes; always enter through the routing endpoint.

orchgraph is a custom C++ fork of Memgraph Community 3.11 that implements clustering, HA, and load-balancing at the C++ level with Raft-based high availability. Three Raft coordinators elect and monitor two data instances; MAIN/REPLICA roles are assigned dynamically by the coordinators, and failover is automatic and native — there is no external HTTP controller.

Layout

Client Connection

Use a routing driver against a coordinator; the neo4j:// scheme auto-routes writes to the current MAIN and reads to replicas, and survives failover:

MEMGRAPH_URI=neo4j://10.53.1.160:7691   # auth=None

Never hardcode a data-instance IP as the write target — the MAIN can move.

Replication & Failover

Single MAIN with asynchronous replication to the replica. The Raft coordinators own cluster membership, role assignment, health monitoring, promotion, and failback natively — no manual role Cypher or curl endpoint is involved. Inspect cluster state over Bolt to any coordinator on port 7691:

SHOW INSTANCES;

Databases on this cluster

The cluster is MG_ENTERPRISE=ON, so CREATE DATABASE gives per-database isolated storage, replicated by the same Raft group — two logical graphs share this one physical cluster, not two clusters:

Select the database explicitly on every session (database=orchkg or database=memgraph) — the routing URI above is the same for both.

Vector Search

The vector_search MAGE module is live on orchgraph:

• CALL vector_search.search(...)
• CALL vector_search.search_edges(...)
• CALL vector_search.show_index_info()

Four indexes are present:

History

The previous cluster — a 3-node Memgraph Community cluster (memgraph-1/2/3) fronted by a custom Python HTTP HA controller (memgraph-ha, :8080, /main /status /failover /failback) plus a standalone memgraph-db — was decommissioned and deleted on 2026-07-10. Both graphs (wiki :Chunk and orchestrata :SemanticChunk) were ported to orchgraph and confirmed serving with a verified parity check before the old cluster was taken offline and deleted.

Backups

orchgraph-backup.service (timer daily at 03:00) runs /opt/orchestrata/scripts/backup_orchgraph.sh --retain 14. It finds the current MAIN, runs CREATE SNAPSHOT; on every database, pulls each newest snapshot with incus file pull to /var/lib/orchestrata/backups/orchgraph/<run>, writes manifest.json with each database's node count (MATCH (n) RETURN count(n)), and prunes to 14 runs. On 2026-10-23 the archive held 14 runs, 4.3 GB. The 2026-10-22 run pulled 213 MB for memgraph and 84 MB for the second database, which is stored under its UUID directory ab169e8f-.... The unit failed on every run from its install until commit 8b2038b92 (2026-10-05); see section 9 of Host services.

Test and dev graphs

These are not part of orchgraph. They are separate containers in the other environments (Incus fleets).

Prod is never a test target. The test suite's graph is orchestrata-test:memgraph.

Related

• Incus fleets
• Memgraph HA cluster build log (historical)
• Neo4j cluster (decommissioned)
• Wiki graph indexer schema
• KG migration record
• Host services